---
格式版本: 2
标题: "Turning Models Into Engines: The AI Factory Era S81715 | GTC San Jose 2026"
原文链接: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/"
发布日期: "2026-09-07"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "Published Time: Mon, 07 Sep 2026 15:17:06 GMT"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 4
发布时间严格候选数量: 1
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-07T23:17:11+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-07T23:16:22+08:00"
入库时间: "2026-09-07T15:18:57.711Z"
来源平台: "固定入口"
搜索渠道: "fixed_url"
搜索词: "https://www.nvidia.com/gtc/"
匹配关键词:
- "Compute tray"
- "Vera Rubin"
- "NVLink"
- "delivery"
- "deployment"
- "performance"
- "latency"
- "bandwidth"
- "GPU"
- "AI"
相关厂家:
- "NVIDIA"
相关专家:
[]
内容类型: "网页"
抓取工具: "Jina Reader"
清洗工具: "Jina Reader Markdown + Defuddle/Readability 正文提取"
原始附件:
[]
AI优质: "是"
AI打分: 81
AI分档: "高置信优质"
AI质检状态: "通过"
AI打分理由: "正文为Mistral AI在GTC 2026的AI Factory演讲实录，主线是Mistral Compute的GB200/GB300机架部署、云原生软件栈、裸金属/托管Kubernetes/SaaS三层服务及2025—2026年建设路径，直接讨论AI Factory和机架级基础设施。来源为NVIDIA官方GTC session页完整转录，属一手会议内容。历史库未见同主题重复稿；正文新增可核验事实：GB200全部投产、GB300软件栈部署、ECODC瑞典数据中心用于Vera Rubin、首个外部客户上架、Koyeb收购等。命中量产/客户/部署通道和生产级深技术通道，未命中硬否决项。"
AI质检模型: "zj-deepseek-v4-flash"
AI质检时间: "2026-09-07T23:19:48+08:00"
AI主题相关性: 18
AI来源权威性: 13
AI新颖性: 15
AI技术细节: 13
AI商业部署信号: 14
AI完整性: 8
AI评分提示词版本: "v17-精简生产版"
AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d"
AI评分知识库版本: "knowledge_base_v1-20260819+runtime.100"
AI评分知识库SHA256: "7fe4568b9de9fc322c105ece08d18efab5b9b86f27dea6eb043b9a4d273bde60"
AI评分知识库检索词: "[\"NVIDIA\",\"Nvidia GTC大会\",\"https://www.nvidia.com/gtc/\",\"GB200\",\"GB300\",\"Vera Rubin\",\"Blackwell\",\"Rubin\",\"NVLink\",\"Compute Tray\",\"RAS\",\"GPU\"]"
AI评分知识库命中: "[{\"id\":\"runtime-f42510f490b4e8271c10e963\",\"title\":\"NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt\",\"sourceType\":\"ai_excellent_article\",\"time\":\"\",\"matchedTerms\":[\"NVIDIA\",\"https://www.nvidia.com/gtc/\",\"GB200\",\"GB300\",\"Vera Rubin\",\"Blackwell\",\"Rubin\",\"NVLink\",\"RAS\",\"GPU\"],\"rank\":-20.93064533391618},{\"id\":\"runtime-dccd09f98e2ca028070793f0\",\"title\":\"https://blogs.oracle.com/cloud-infrastructure/oci-achieves-nvidia-exemplar-cloud\",\"sourceType\":\"ai_excellent_article\",\"time\":\"\",\"matchedTerms\":[\"NVIDIA\",\"GB200\",\"GB300\",\"Blackwell\",\"NVLink\",\"RAS\",\"GPU\"],\"rank\":-18.19676685780169},{\"id\":\"july-correct-0020\",\"title\":\"NVIDIA Vera Rubin：引領代理 AI 的時代\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"NVIDIA\",\"Vera Rubin\",\"Blackwell\",\"Rubin\",\"NVLink\",\"RAS\",\"GPU\"],\"rank\":-17.245230523944805},{\"id\":\"july-correct-0104\",\"title\":\"NVIDIA Vera Rubin 提升每瓦性能，为全球合作伙伴实现最低 Token 成本\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"NVIDIA\",\"GB200\",\"Vera Rubin\",\"Blackwell\",\"Rubin\",\"NVLink\",\"RAS\",\"GPU\"],\"rank\":-17.094245285367755},{\"id\":\"july-correct-0095\",\"title\":\"Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"NVIDIA\",\"Vera Rubin\",\"Blackwell\",\"Rubin\",\"NVLink\",\"RAS\",\"GPU\"],\"rank\":-16.474365960390056}]"
采集批次: "2026年9月7日23点15分57秒"
采集批次ID: "20260907-231557-722"
去重键: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715"
---
Title: Turning Models Into Engines: The AI Factory Era S81715 | GTC San Jose 2026
URL Source: https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/
Published Time: Mon, 07 Sep 2026 15:17:06 GMT
Markdown Content:
Visit your regional NVIDIA website for local content, pricing, and where to buy partners specific to your country.
[Continue](https://www.nvidia.com/)
Upcoming: Explore other GTCs: [**GTC Berlin** October 20–22](https://www.nvidia.com/en-eu/gtc/) |
[**GTC Washington, D.C.** November 30-December 3](https://www.nvidia.com/gtc/dc/)
| [**GTC 2027** March 15–18](https://www.nvidia.com/gtc/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [](https://www.nvidia.com/en-us/account/)
* [Log In](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)[LogOut](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
PLATFORMS
other links
[](https://www.nvidia.com/gtc/)
Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
[Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/# "Menu")
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/# "Menu")
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81715/#)
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
Loading
Play
00:00
Play
Seek 10 seconds backwards
Seek 10 seconds forward
00:00 / 00:00
Mute
Press TAB to open volume control Use the arrows to control the volume
Turn on Picture in picture
Show Full screen
Transcript Powered by AI
X
X
00:09
So good afternoon CTC. Hello everyone.
00:12
So we'll present you how and why a frontier AI startup decided to
00:19
expand to the AI factory business.
00:25
So first, a short update on Mistral AI, which was funded
00:30
by three French mousquetaires, Arthur, Guillaume, and Timothée,
00:34
less than three years ago.
00:36
We are now more than 600 employees.
00:39
No, no, no, no. You are not up to date, Gautier.
00:42
Yeah, so we are today, I'm sure, more than 900 employees, thanks
00:46
to the Talent Acquisition team.
00:49
And there is more than 13 nationalities represented.
00:54
And when they are not working remotely, you can find them
00:58
in one of our office, like the new flagship at quarter
01:03
in Paris, or one of the two office we have in the bay,
01:07
or an office in Singapore, London.
01:10
I cannot name them all.
01:12
There is something common in all the Mistral employees that we
01:17
are really driven by the mission.
01:20
Yeah, and regarding this mission, it's really to bring AI as
01:23
a tool and something which is very specific on a per-customer basis.
01:28
So the purpose, and as is stated in the slide, is to empower
01:31
customers and partners, by the way, with the use of AI for on purpose.
01:36
It's not something that is just doing AI because it's
01:39
nice, because it's hype, but it's really something that
01:41
we can qualify, that we can formalize, and where we are able
01:45
to get, I would say, some kind of
01:47
What is the value behind?
01:49
So it's really, really important for all the team within Mistral
01:51
AI that we are able to go through that and provide outcome
01:56
with KPI, ROI and things like that.
02:00
It's also very important that we are not, I would say,
02:03
a service company.
02:05
We are building AI, we are building frontier AI models,
02:08
so it's really the core of the company, and also we are
02:11
building tools which are here ready to help using models, developing
02:15
new models, and so on and so forth.
02:17
So it really is a global approach we have within Mistral AI.
02:23
So regarding this aspect, we are already trying to go
02:26
through three steps, three main steps, let's say, on this approach.
02:30
It's really to build frontier AI model, but I would say specific
02:34
models on a per-case basis.
02:36
So really something which is focusing what customers
02:40
are expecting or, I would say,
02:42
Are we adapt to the given market?
02:44
It's also to bring this AI, not only on something theoretical,
02:49
but in something which is practical that can be used
02:51
on a daily basis through agents, through application, through API.
02:55
So something that people and users can rely on on a daily basis,
02:59
whatever the jobs, whatever the context, it's very important to us.
03:03
And the last point is that we can, I would say,
03:06
It's stated as realized, but it's that we need to realize
03:09
value, realize something which is an incentive, which is
03:13
something new for the company that brings something as an
03:16
addition to the company.
03:19
We have, I would say, very standard stuff like improve continuously.
03:23
I would say, especially with AI, we will not say, OK, nothing's moving.
03:27
We have done something and so we rest.
03:30
So we need to improve over the time and so it's long run,
03:34
it's not something that's one-shot.
03:36
Second point, as Gautier mentions, we are a French company, so
03:42
we are European-based, and governance, security, law
03:46
management is very important, so governance, everything's taking
03:50
into account this type of things and customers are expecting that.
03:54
The last point is redeploying everywhere.
03:57
So as I mentioned, we want to facilitate the use of AI and where
04:02
we deploy is very, very important.
04:03
So we are able to deploy on edge, on premise, on cloud and so far.
04:09
But we have no control on the compute resources.
04:14
It's lead us to improve on that aspect and create an
04:18
infrastructure which really purpose will for AI workloads.
04:21
So for all internal teams, for all internal workloads, have something
04:25
which is named Mistral Compute, which is really something where we
04:30
have control and we have end-to-end
04:35
solution for AI workloads.
04:39
On this point, I mean, Mistral AI is really based on what
04:45
is stated on the slide, which is many, many, many people
04:48
can bring compute resources.
04:55
However, very few are able to provide AI at scale and with
05:00
the expected level of reliability, especially for AI workloads.
05:05
So it's roughly the same as on HPC market, but with an
05:09
higher degree of complexity.
05:12
So what we try to build is really what NVIDIA is
05:14
mentioning as AI factory.
05:16
So even Jensen mentioned again, AI factory during the speech.
05:19
It's very important to us to build something which is
05:22
from AI builders by AI builders.
05:24
So it's really something targeting people doing AI.
05:29
So the overall infrastructure is really targeting the time to
05:34
train and the time to infer as well as the reliability of the solution.
05:39
You see, my accent and my English is not perfect.
05:42
I am French, so... The second point is really the sovereignty aspect.
05:46
As I already mentioned, this is very, very crucial.
05:51
We state in the slide that it's all destiny and control
05:55
all destiny, Mistral AI destiny.
05:57
If we have no compute resources, if we have no storage, if
06:00
we have no network,
06:02
I mean, we are dead.
06:04
In any AI companies, if you do not have such type of resources,
06:08
you cannot perform.
06:09
You just lose a race.
06:11
The race is already very challenging.
06:12
So if we do not have compute resources, we are just, we
06:16
will not exist anymore.
06:17
So it's very important to us.
06:18
That's really the two drivers for Mistral Compute.
06:24
So now, what are the key building blocks of Mistral Compute?
06:28
Of course, data centers, infrastructure.
06:30
I will not go through that. I mean, it's quite obvious
06:32
that we need such type of things, even if we can spend
06:34
half an hour or maybe more on that.
06:37
The main aspect is really what type of offers we are able to propose.
06:41
So first one is bar metal.
06:42
When I say bar metal, it's really bar metal.
06:44
It's not something like bare metal as a service.
06:46
It's really a pure bar metal offers where customers are...
06:50
Purchasing hardware.
06:52
And we just provide it through data center, through co-locations,
06:56
and give them to them.
06:58
It's mainly focusing hyperscaler because it's
07:00
a very dedicated infrastructure.
07:02
We have, I would say, fully access to the hardware, to
07:04
all controllers, to all the networks, so it's not something
07:06
that you can share easily.
07:08
It's too complex.
07:09
The second offer, which is really the mainstream offer
07:12
for Mistral Compute, is what we call Managed Orchestrator.
07:16
Roughly speaking, it's a managed Kubernetes.
07:18
I mean, Kubernetes is a standard, de facto standard, for AI.
07:23
So it's really bringing Kubernetes to our customer.
07:28
It's for AI labs, for industrials, for small and medium enterprise.
07:31
So it's really broad, or even hyperscaler, as they wish.
07:35
It's very, very, I would say, a generic offer we can go through.
07:39
What is very important also to us is that we have in our
07:42
DNA, I mean, in Mistral DNA, we have open source model, open
07:45
weight models, open source is key.
07:48
I mean, it's not just, okay, let's pick some stuff, let's
07:51
make some stuff aside.
07:53
It's really building something as open source,
07:56
open source as a foundation.
07:58
And also use standards.
07:59
So we can rely on many open source projects.
08:01
If we are not going to standards, it brings some complexity
08:05
as also some kind of vendor lock.
08:08
One stuff we really want to avoid is to go in this direction.
08:11
We want to have something open where people can use and rely on
08:14
technology which are, I would say, back on communities and where we
08:18
are providing stuff to communities.
08:21
Deeper diving on that.
08:23
So it's a little bit more technical part here.
08:25
It's really what are the technical components we are going through
08:28
here, especially in open source for Mistral Compute.
08:32
First one is.
08:33
Relying on Kubernetes is not as trivial as it looks.
08:37
It's really relying on key components and not make derivations
08:42
from the standard.
08:43
And at scale, we have a lot of resources, not completely trivial.
08:47
So the first element we were going through is relying on
08:51
the cluster API, CAPI.
08:54
Why we choose to go this way is we are able to put a lot
08:58
of things with Kubernetes.
08:59
People using it know that perfectly well.
09:02
It's insane how we can do stuff.
09:04
So we look at which provider we can take to manage the infrastructure.
09:11
So really, with MetalCube, it's all the stuff regarding
09:15
the hardware deployment.
09:17
So the compute node deployment, mainly.
09:19
And how we are able to integrate them directly within Kubernetes.
09:23
Normally, it's two separate worlds.
09:25
There is one world which is the hardware, and there is
09:27
the other world which is the logical software side.
09:29
With MetalCube, we are able to handle hardware resources
09:33
as they were Kubernetes resources, which make things very, very,
09:37
very convenient and easy to manage on a daily basis.
09:41
The second element is multi-tenancy.
09:45
Okay, we will build very large cluster, as we have today.
09:48
We need to slice them.
09:49
I mean, we will not have one big cluster all the time,
09:52
so we need to make multi-tenants.
09:54
And for that, we are relying on Camagy, which is also an
09:57
open source project, not a big surprise here.
10:00
Relying also on the company on the back, which is Clastics.
10:03
It's an Italian company developing Camagy.
10:06
So here, the purpose is really to make Kubernetes inside Kubernetes.
10:11
So it's really hosted control plane, so it's
10:14
really hosted Kubernetes.
10:16
The main interest that customers have a full Kubernetes
10:19
infrastructure available.
10:21
So it's not a kind of some part of, so you have the full
10:24
access and the full control of your Kubernetes instance with
10:29
a very high level of security for segregation between the different
10:32
tenants and so different customers.
10:35
Then another topic is, OK, we have plenty of resources to manage.
10:39
We have a lot of things to handle.
10:42
How can we handle multi-clusters, thousands
10:46
of resources in each cluster?
10:49
Again, in an easy way, how can we sort that out, I would
10:53
say, through open source?
10:55
We looked for open source projects and we didn't find any.
10:58
To be honest, it's quite specific.
11:00
So we start developing a compute operator that allows us to
11:05
handle all resources, like one is maybe a little bit
11:08
too much, let's do some marketing.
11:10
But really, it's, okay, how can I handle the amount of
11:13
resources on a per cluster, on a
11:16
And have all the status of all resources and all the
11:19
services that are running on, like if it was very few.
11:23
It's something we expose also to customers through API,
11:26
through identity provider, just for security.
11:31
Then another very, very tricky topic is how do we have a full view
11:38
on what happened on the system?
11:40
It's quite complex to have no blind spots regarding infrastructure,
11:45
regarding network, regarding storage, regarding compute,
11:48
regarding services that we are running on top of compute,
11:51
and so on and so forth.
11:52
So we were building an observability stack
11:55
with usual suspect.
11:57
I mean, we didn't try to reinvent the wheel.
11:59
We don't try to make something.
12:01
Fancy. We just try to make something that works.
12:03
So we rely on Grafana with Loki, which is also from Grafana
12:07
for the logs, Mimir for the events, Alloy as a gateway,
12:13
and all of that through the OpenTelemetry, through OTEL.
12:17
Why OTEL? Because, again, it's a standard, but also
12:20
because we can have information from hardware, from software,
12:23
from services, which are exactly from the same format and with
12:27
the same way to handle them.
12:31
As an addition, one tricky stuff was to be able to
12:35
deploy such a stack.
12:37
So OK, we have one tenant.
12:39
It's quite easy. We have two tenants.
12:40
Let's do by hand.
12:41
We have 10 tenants.
12:42
OK, let's try and symbol, or things like that, or end charts.
12:47
When we have hundreds of tenants with very few deltas between each
12:50
of them, it starts to be painful.
12:52
And Sveltos, again, an open source project, really helps
12:56
us to build something that is easy to deploy, easy to manage.
13:01
I mean, to control what is the status of the different
13:03
resources, the reconciliation loops, and so on and so forth.
13:07
So Sveltos is really here to help us to
13:11
Spread the services and especially the monitoring, the observability
13:16
elements over the cluster and over the tenants.
13:21
Then we build Kubernetes.
13:24
OK, good.
13:27
Should we provide some extra service?
13:28
We do not plan to provide hundreds of services, maybe
13:31
later on, but at least today, this was not our main focus.
13:34
Where we want to go is really to provide some
13:37
key services where customers are feeling, OK, I am good.
13:41
All the infrastructure is OK.
13:44
I have my storage.
13:45
I have my network.
13:47
I am able to submit jobs, whatever they are, through
13:49
Kubernetes or to Slurm.
13:51
So we are providing all of those add-ons like part of the
13:55
product, part of the solutions and bring something like an out-of-box
14:01
solution for people relying on the managed orchestration,
14:04
managed Kubernetes solution.
14:06
And again, to deploy them, we are relying on Sveltos
14:09
because it's quite convenient for, I would say, monitoring services,
14:14
it's exactly the same for add-ons.
14:15
So we can spread them through Sveltos the same way.
14:18
So again, not reinventing the wheel, using the same
14:21
software, less issue at the end.
14:25
So moving back to the, I would say, building blocks
14:28
and offers, so BarMetal,
14:31
Managed Orchestrations.
14:33
And so, okay, we are an AI CHOP.
14:35
So we are providing AI systems.
14:38
We are providing AI tools for people to use them.
14:41
So here, the purpose is, okay, on top of Mistral Compute,
14:44
we provide Mistral AI.
14:46
It could be others, but we provide Mistral AI skills,
14:51
knowledge, tools, models.
14:53
on top of Mistral Compute through a SaaS model.
14:56
So it's really a service approach.
14:58
So people will not care about the infrastructure,
15:01
the belonging infrastructure.
15:03
It could be something other than compute, but it could be compute.
15:05
So it's sovereign, it's something under or for control.
15:08
So data are under control, compute resources are under
15:11
control, and so on and so forth.
15:15
As transversal components, we have, I would say, again,
15:18
things which are very common for, I would say, any solution provider.
15:23
So, we need to manage support.
15:25
So, we need to have people handling support requests,
15:28
and we have, for sure.
15:29
But we work with our AI team to enhance that and to use
15:33
AI to, I would say,
15:35
Is the support path not, as we said, just to put a bot?
15:39
That's not the idea.
15:40
I mean, we want to have people, but to help people,
15:42
okay, that's an issue.
15:44
What are the potentials and what can I do from a
15:48
technical perspective?
15:48
So I have a kind of expert aside, 24 hours, 7 days a week.
15:55
So it's very important to go through that.
15:57
Lifecycle management is also something we can really enhance
15:59
through AI because we want to keep the system up and
16:02
running so we can use AI to understand corner case and finally
16:06
make corner case not anymore corner case, but something that
16:09
can be sorted out very quickly.
16:11
And of course, security, which I can spend one hour or more
16:15
just on the security aspect.
16:17
Last point on the offer is that, okay, we are building
16:19
a lot of stuff, but it's only thanks to Mistral AI team,
16:24
Mistral Compute team also, but I mean, we are a global team.
16:28
And it's very, very in, again, in the DNA of Mistral that
16:33
we are working through two teams and whatever it is, infrastructure,
16:37
whatever it is, even procurement.
16:39
We have people from procurement in the room, exactly, and
16:44
technicals from AI, from software stack, from data center, we
16:48
are really working together.
16:49
It's quite impressive to be honest, I have joined Mistral
16:51
not so far, and we are really working all together to reach
16:55
some goal for the customers.
16:58
So now we will go through the timeline.
17:01
So Gauthier, if you can explain our journey in a few time.
17:04
GAUTHIER DURAND- Yeah, so we'll go through the journey
17:08
of Mistral Compute.
17:10
Actually, we really started in April, so last year, less
17:15
than a year ago, with the idea that if we are able to
17:19
provide resources, computes, GPUs to our internal team, we must
17:25
be able to provide it also to other AI startups and other customers.
17:31
So in April, we decided to accelerate the acquisition
17:37
of more chips, more data centers.
17:40
Then in June, at GTC Paris,
17:45
Jameson, Archer, and President Macron announced Mistral Compute.
17:52
And then in July, we received our first GB200 racks in our data
18:00
center, and we also started the design of our own software stack.
18:06
After a very busy summer in September, we onboarded the
18:12
first employees of Mistral Compute to build the teams.
18:16
Georg is one of them.
18:19
And we were still receiving more and more GB200.
18:25
Then before Christmas, in December, we received the first
18:28
GB300, next generation of systems.
18:32
We also have been able to deploy our first version of the software
18:37
stack on a few GB300 racks.
18:43
And to start 2026, we also integrate the GB300 racks
18:49
in our cluster with our own software stack.
18:55
And last month we announced a new partnership with ECODC
19:00
and thanks to them we will have a new data center in
19:04
Sweden and also we finished the deployment of all the GB200 and we
19:11
have all the GB200 in production.
19:14
And also for EcoDC, so it will be a data center that
19:19
we'll use for the next generation of GPUs, which are the VeraRuby.
19:27
And today we are onboarding our first external
19:31
customer on our systems.
19:37
I would like to also to tell you more about the lesson
19:42
learned during the bring up of this system, or first bring
19:45
up at Mistral Compute.
19:47
So first we have to understand that for this large-scale
19:51
project, a delivery takes months, and then the installation
19:58
and the validation of the system.
20:01
Last weeks, so all these steps are done by OEM and NVIS, and when
20:10
they're finished with validation, there is a handover phase.
20:13
Roughly, they give us the keys of the system, and then
20:20
we'll do our own validation.
20:21
So at that stage, we will deploy our software stack on the
20:26
system and we run on validation.
20:30
So for validation, we are using mostly nickel and nemotron.
20:36
I always use nemotron across many, many racks to stress the system.
20:41
And as soon as I thought the system was stable, I gave the system
20:49
to the internal team, Mr. AI.
20:52
And I remember an evening, so I was in a call with them.
20:56
And they asked me, OK, do you think the system is stable for us?
21:00
Can we start to run jobs on them?
21:03
And obviously, I told them, yes, of course.
21:05
I mean, I'm running a for days.
21:08
There was no hardware failure.
21:10
No issue.
21:11
Let's go. And I guess you know what would be the next step.
21:15
They launched the job, and after a few minutes, the system crashed.
21:20
What did you do?
21:21
So what I did, of course, I banned the user.
21:24
And then the system became stable again.
21:28
The easy way.
21:30
But for real.
21:31
For real, what we did, of course, we integrated their jobs.
21:36
Their code base into our validation and stress suite.
21:40
So we continue to stress the system with their own code,
21:45
so we are more hardware filler.
21:48
And at the end, they have been able now to use the system without
21:54
crashing the hardware too often.
21:57
So the lesson learned from this is really that
22:05
Our science team, the people who are developing the software
22:09
to train the model, are pushing the hardware to their limits and they
22:14
are also, with the new techniques, they will stress this hardware in a
22:19
different way than they did before.
22:22
So it's very important for us to always update our validation
22:27
and stress suite to integrate everything we are doing.
22:33
To ensure that the system will be stable.
22:36
I would also thanks NVIDIA, because they have been very
22:39
crucial, the support from NVIDIA has been crucial doing all these
22:43
steps, especially with the NVLink troubleshooting, also to understand
22:46
why we had some individual.
22:47
ALAIN VONGSOUVANHOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUMOUM
22:56
And even if so, I mentioned that it's very important for
23:00
us to have, to integrate our own code-based validation.
23:08
I would emphasize that all the steps before installation
23:12
validation by NVIS or by us, it's very important and we
23:17
cannot skip any of them.
23:19
But at the end, again, real use case, real workload are
23:28
very important keys for that.
23:34
Okay, so we are roughly at the end of the presentation, so
23:38
let's try to make some takeaways.
23:41
So the first one is that we are already trying to provide
23:45
maximum performance in terms of on the latest GPU.
23:49
So the purpose is to bring something for our internal
23:52
team which are at the maximum of what the chips can provide,
23:57
chips, network, and storage.
23:58
So that's really the first element of what we are trying
24:02
to seek with Mistral Compute.
24:03
Thank you for watching.
24:04
The second one is, as Gautier mentioned, we are developing
24:09
stuff which are, I would say, easy loading.
24:13
Whatever it is on compute, on storage, on the network,
24:15
it's burning, stressing all the components at maximum.
24:20
So it's very, very, very tricky to ensure that we have something
24:24
which is reliable enough over time and at scale.
24:28
So it's what we are trying to build on, I would say, on our journey.
24:32
The other point is that we build different type of offer
24:36
because we understand that there is not one fitting all.
24:39
So there are customers that are willing to have full control
24:41
on the infrastructure.
24:42
Okay, there are customers that wish to have, say, managed
24:46
Kubernetes and have part of the control, and some which do
24:49
not care about the infrastructure and just wish to be able to run AI.
24:53
That's, I would say, our core business, that running AI
24:57
is what we are seeking.
24:59
And we are in a very, very specific position where, OK, we have smart
25:04
people, not to say more than smart,
25:08
and we have very huge benefits relying on them on a daily basis
25:12
for the design of the solution and for the operation of the solution.
25:16
So it's...
25:18
For short, it's what works for our AI team.
25:21
We finally don't catch why it may not work for our customers.
25:25
So it's really the purpose of building that and
25:28
opening that to outside.
25:30
So we can keep it inside, of course, and we will rely on it.
25:34
But finally, what we do could be some benefits for our customer.
25:38
And it's a really unique combination of
25:41
What AI companies can bring and what people doing infrastructure
25:45
and compute resources can do.
25:47
So this combination is quite, yeah, specific, let's say.
25:51
And if we just get back to one of the first slide, the title was
25:56
Many Can Provide Compute Resources, Few Can Run AI Workloads.
26:00
We can run AI workloads and we can give proof on that aspect.
26:04
So it's really, really where we want to go and where we
26:06
are going until roughly one year.
26:10
If you want to continue to have, I would say, some information
26:13
about Mistral and what we are doing, there is two other
26:17
presentations during GTC, one with our CTO and co-founder,
26:21
Timothy, and another with our CEO and co-founder, Arthur.
26:25
And also, feel free to join us on the booth and
26:28
to have some discussion.
26:29
We have a booth at the outside, we have a booth at the inside.
26:33
Again, we are really here to have some discussion and
26:35
to also learn from you all.
26:38
So we are done.
26:40
Thanks a lot. We hope it has been quite interesting.
26:42
Let's enjoy GTC.
26:45
And if you have any question, we are trying to answer.
26:50
At least we're trying. Do your best.
26:55
Any questions?
27:02
Question regarding reliability, which is, let's say you lose,
27:06
at some point you lose a node in your infrastructure and
27:12
one of your customers had a several weeks job running on it.
27:14
Do you provide something to handle this or is it their
27:19
own responsibility to be able to restart something, whatever?
27:22
Are you able to move a container to another node in real time or not?
27:28
You still hear me? Yeah, okay.
27:31
So it really depends on what type of offer.
27:33
I mean, for BarMetal, it's quite obvious, out of the scope.
27:35
For managed Kubernetes, no.
27:37
It's really on the customer responsibility to under this path.
27:40
If it's true Mistral AI stack, yes, it's something we can work on and
27:44
to have, obviously, solutions that ensure that we are checkpointing
27:48
and be able to restart application without losing weeks, days,
27:51
or months of application.
27:53
So yes, but really depends on which level you are going to.
27:56
For managed Kubernetes, as of today, it's retargeting
27:59
managed Kubernetes, pure managed Kubernetes.
28:01
So we view, I would say, extra services on top.
28:04
It's something we are looking also to do checkpointing,
28:06
system checkpointing, and things like that.
28:09
But it's longer term topics.
28:10
I mean, we already have some work to do.
28:15
I wonder how, so how onerous is it sort of staying in sync with the
28:22
EU legislation on AI governance?
28:25
Is it something you find is a lot of overhead or no?
28:28
I'm curious. Another one.
28:31
From my perspective, it could be tricky.
28:35
Yeah, let's be short.
28:36
Yes, it could be tricky.
28:37
Is it too much? I don't think so.
28:39
It's much more depending on which verticals you are looking at.
28:42
If you are looking for health science, I mean,
28:44
it's maybe not enough.
28:45
If you are looking for commodities, potentially it's far too much.
28:48
So putting the good, I would say, the good level, the good
28:51
trade-off is quite tricky.
28:52
And especially for people, sorry to say that, but especially for people
28:55
who do not master the technology.
28:57
It's quite tricky to make the good balance between, okay,
29:00
making very, very huge protection for people, for citizens, and
29:04
so on and so forth, and keeping in mind that you have to run business.
29:08
So today, I think we try to find a good balance.
29:11
It's not easy, to be honest, not only for AI, but also for security
29:15
and for regulation, organic security, with the Cyber Resilience
29:18
Act, for example, in Europe, which brings a lot of constraints.
29:21
But again, it's very dependent on the type of customer you
29:24
have in front of you.
29:24
So maybe we need to find a better way to adapt regulation,
29:28
not only a general-purpose regulation, but something which is
29:31
much more specific, like it exists for defense and things like that.
29:35
But again, I'm sure it's very tricky, so it's not my world,
29:40
but I think it's not too much.
29:42
Again, we need to have some control on that.
29:44
To me, it's quite important to have it.
29:46
My question has to do with the way how you are mixing
29:50
the AI services, AI modeling.
29:54
So I own an SCP in the United States.
29:58
Are you aligning and having a network of data centers?
30:03
Or are you actually buying data centers?
30:05
We are doing collocation in data centers.
30:07
So we have room in data centers as of today.
30:09
Let's see what will be the future.
30:11
But as of today, yes, we have close.
30:13
We have multiple data centers close to Paris right now and
30:16
the other ones that Gauthier mentioned in Sweden.
30:20
But it's not a network of data center.
30:22
As of today, they are operating as standalone data center.
30:25
We are looking also to make some kind, especially for applications,
30:29
to be sure that we have, in case of disaster, disaster recovery process
30:33
over data center, like geographies, like cloud providers are offering.
30:37
It's not the case as of today.
30:38
It's quite tricky, especially if you want to run jobs across
30:42
multi-geographies, which is something quite complex in terms
30:45
of latency, network bandwidth, and so on and so forth, and technology.
30:49
It's also something people, at least on our side,
30:51
we have, let's say, one technology per data center,
30:55
one generation by data center.
30:57
If you have to cross by data center, potentially you have
31:00
ARM on one side with H100, on the other it will be Vera Rubin.
31:04
It's not completely trivial to handle jobs that are able
31:07
to run perfectly well on the different infrastructure.
31:10
I don't know if it answers your point, but... It does.
31:12
There's so much more to talk about.
31:14
Yeah, exactly.
31:16
Thank you.
31:19
I don't know if you want to add something, Gauthier?
31:21
No, I'm good.
31:23
You alluded to some services that you're going to add
31:26
into Mistral Cloud.
31:27
Can you elaborate a little bit about your idea, like
31:30
long-term vision for Mistral Cloud?
31:32
That's a very, very good question.
31:34
I don't know if maybe you know that we have acquired
31:37
a company very recently named Koyeb that is really focusing
31:41
on sandboxing and things like that.
31:43
So we have some stuff in mind, to be honest.
31:46
So today we are focusing on this.
31:49
It's a slow set of services, but we want to provide them well
31:52
and to be sure that they are really fulfilling what we have in mind and
31:55
what our customers are expecting.
31:57
So today we are not targeting to offer a lot of services,
32:00
but we will extend.
32:01
Thanks to Coyab and thanks to other stuff and relationships
32:04
and discussions with customers.
32:05
And it's where the close relationship with customers
32:07
is very important.
32:09
We don't just want to make yet another stuff.
32:11
So a copycat of others, we want to really focus on something.
32:15
One example is database.
32:17
For AI, we need database.
32:18
So we will provide database services, but it's very embedded
32:21
with the AI offers we are building.
32:23
So it's not core services.
32:28
Thank you so much. First of all, sorry.
32:30
First of all, thanks beaucoup for the great talk.
32:34
De rien. Merci.
32:35
I have a question regarding the cost of inference.
32:39
I'm not sure what you think is the most...
32:42
It's important to lower the cost of inference, especially
32:44
in today's skin talk.
32:47
Do you agree or do you think it's something that we need
32:50
to really care about it?
32:52
Something we should really care about is the model design
32:55
or the hardware design or combine them together.
32:59
You want to take it or I take it?
33:01
You can take it. I can take it.
33:02
Okay.
33:05
At first, I'm not an AI expert.
33:06
So I will be sharp on that.
33:08
I know technology.
33:09
I know infrastructure. I know part of software.
33:12
My feeling, and it would be really a feeling, it's not
33:15
a statement from Mistral or something like that.
33:17
So yes, we need to care.
33:19
That's the first point.
33:20
So again, we need to adapt technology on the different
33:23
evolutions of, I would say, training inference.
33:26
I mean, whatever it is for training or inference or pre-training
33:29
or post-training.
33:29
I mean, we need to adapt the hardware and especially the
33:32
software on these guys.
33:34
And one step which is very important, again, to me, and it's
33:37
a discussion we have internally, is to optimize resources.
33:41
And when I say optimize resources, it's not only, OK, pick the
33:44
right hardware, OK?
33:46
I can make an advertisement for NVIDIA with a GTC for
33:49
sure, but the point is also to look at a GB200 or GB300.
33:56
We have a huge CPU, we have a huge GPU, we have a huge network, and
34:00
when you look at how applications are using such resources,
34:03
it's not 100 over the time.
34:05
So we need to place different types of workloads and make
34:08
a kind of Tetris.
34:10
I think Tetris is a very, very good example to enhance
34:13
the scheduling of resources and to reduce the granularity
34:18
of the resources you are using.
34:20
And then there is, okay, what type of resources can better
34:23
fit my type of workload?
34:24
So maybe it's GPU, maybe it's LPU, as Jensen announced.
34:28
It can be different types of technologies that
34:30
you should combine.
34:31
You have to combine technologies.
34:33
I'm 100% sure of that.
34:35
I hope it answers your point.
34:37
Again, it's very short.
34:39
Thank you so much. You're welcome.
34:42
Hey, how do you do billing for your tenants?
34:46
Excellent question.
34:48
That's a tricky one, to be honest.
34:52
So today for all tenants, we have labels inside Kubernetes,
34:55
so we know which resources are belonging to which tenants.
34:58
That's the entry point to be able to track down what
35:01
type of resources.
35:02
We are enabling Kubernetes accounting to be able to track
35:06
every resources, and also we are looking for some open source tools.
35:11
To make billing, but today we are not sure that we rely
35:14
on them or if we are developing something on our own to make
35:17
a very, very accurate billing.
35:19
Today, again, to be completely transparent on that, we are
35:22
managing at the granularity of a compute tray.
35:24
So it's quite easy.
35:25
I mean, it's not a big deal to handle it.
35:27
When we will really slice GPU, like doing MIG or things
35:31
like that, it will start to be a little bit more tricky.
35:34
So it's work in progress.
35:35
To be honest, it's not, it's, I would say, half
35:38
done and half work to do.
35:41
How does Mistral AI as a business help you to kind
35:44
of build a better cloud?
35:46
Is it, so you talked about like stress testing and ability
35:49
to kind of, they submit jobs so you know how the cloud
35:53
really operates, but there's any other stories in which the
35:56
work from research really help you to just to build a better cloud?
36:00
So I would say it's mainly for the, I mean, validation testing,
36:05
but it's also to understand what we must have in our data center to
36:11
accept different type of workload.
36:13
So what will be the quantity of GPUs, CPUs, LPUs, or other chips.
36:19
Or storage. Storage also.
36:24
It's key also to understand what we need today, what we need tomorrow,
36:28
what they will need tomorrow, and what the market will need.
36:33
And there is also, and I try to mention that very quickly in the
36:36
presentation, it's that we are also working with our internal teams,
36:40
which are really used to work with customers, so end customers, to
36:44
enhance, for example, the support.
36:47
So what is the way that we can go to the root cause of something?
36:51
So you can do by hand.
36:52
We have experts. I mean, the question is not we.
36:54
We have very, very strong experts in the teams.
36:58
But again, we do not have them available 100% of the time.
37:03
And there are many steps that you can finally analysis and use AI
37:06
to speed up this type of analysis.
37:08
It's the same for everything regarding lifecycle management.
37:11
So when we have failures, we are able to track down
37:14
and to anticipate, predict.
37:15
It's a kind of prediction.
37:17
Predictive maintenance is something also we work in with the team
37:20
because we have customers doing predictive maintenance, like ASML.
37:23
I mean, ASML needs predictive maintenance, so we work in
37:26
the same mindset as with ASML for the Mistral Compute team.
37:31
So it's really, I mean, I have no strong example, like,
37:33
okay, we sort things that we handle days to sort it
37:36
out and we take just a few seconds, like failure with a science team.
37:40
But, I mean, it's exactly the way we want to perform.
37:42
And to me, it's really the benefits working within AI teams.
37:46
That's really good.
37:50
I'm with Proximus Luxemburg, we are partner of Mistral and my question
37:55
was, when it comes to this Mistral Compute platform, is that something
38:00
you consider also, you know, giving to partner to operate it?
38:04
I mean, could I as a partner purchase this whole pack,
38:07
you know, Mistral Compute with, of course, the frontier
38:10
model and everything, I operate it on behalf of our customers
38:13
and you guys to the Altru3 support in a classical operator model.
38:18
Okay, are you from our strategic team or not at all?
38:23
No, but... Your question is great, so thank you for it,
38:26
because it's also something.
38:28
So today we are providing that through our infrastructure,
38:31
and it's something we are thinking about.
38:33
We have some customers asking about, they have data centers,
38:37
they have infrastructure, but they don't want to bother.
38:39
Let's say we're managing it, having a software stack which is very
38:43
efficient and be able to handle it.
38:45
So I will not say yes or no.
38:47
It's something we are really thinking about to have partners
38:50
that is a kind of some extent a kind of private Mistral
38:53
Compute entities that is completely under customer control.
38:58
But we have our software stack and also benefit of
39:00
our expertise and personally of services on top of it.
39:04
It's something under discussion.
39:05
It's not something very confidential, but it's something
39:08
we are looking at.
39:10
So if you are interested, go to us and let's have some discussion
39:13
because I think it's something we push to, I would say, to
39:16
Timothy, to Arthur, and Guillaume.
39:19
So it's really, really important.
39:21
Thank you for this.
39:23
Thank you, everyone.
39:24
Thank you. First speaker.
39:26
Have a good day. Thank you very much.
# Turning Models Into Engines: The AI Factory Era
Gauthier Delerce,HPC Lead,Mistral AI
Jean-Olivier Gerphagnon,Software Architect,Mistral AI
Rate Now
Step inside Mistral’s journey to industrial-scale AI as we share lessons learned and the principles guiding our approach in creating Mistral Compute. Built on four key pillars—open-source foundations, security by design, sustainability at scale, and digital sovereignty—we’ll share how our approach unlocks bare-metal GPU performance, why that matters in building trusted, European-hosted future-ready AI systems, and how we’re bringing collaboration to the next level by giving innovators access to the same infrastructure powering our models.
### Learn More About This Topic
Share
Favorite
Add to list
PDF
Events & Trainings:GTC San Jose
Date:March 2026
Industry:Cloud Services
Topic:Data Center / Cloud - Software-Defined Data Center
Level:General Interest
Language:English
NVIDIA technology:Cloud / Data Center GPU,Grace CPU, Infiniband Networking,Ethernet Networking,Interconnect Networking,Blackwell
Region:
Company Information
* [About Us](https://www.nvidia.com/en-us/about-nvidia/)
* [Investors](https://investor.nvidia.com/home/default.aspx)
* [Venture Capital (NVentures)](https://www.nvidia.com/en-us/startups/nventures/)
* [NVIDIA Foundation](https://www.nvidia.com/en-us/foundation/)
* [Research](https://www.nvidia.com/en-us/research/)
* [Corporate Sustainability](https://www.nvidia.com/en-us/sustainability/)
* [Technologies](https://www.nvidia.com/en-us/technologies/)
* [Careers](https://www.nvidia.com/en-us/about-nvidia/careers/)
News and Events
* [Newsroom](https://nvidianews.nvidia.com/)
* [Company Blog](https://blogs.nvidia.com/)
* [Technical Blog](https://developer.nvidia.com/blog/)
* [Webinars](https://www.nvidia.com/en-us/about-nvidia/webinar-portal/)
* [Stay Informed](https://www.nvidia.com/en-us/preferences/email-signup/)
* [Events Calendar](https://www.nvidia.com/en-us/events/)
* [GTC AI Conference](https://www.nvidia.com/gtc/events/)
* [NVIDIA On-Demand](https://www.nvidia.com/en-us/on-demand/)
Popular Links
* [Developers](https://developer.nvidia.com/)
* [Partners](https://www.nvidia.com/en-us/about-nvidia/partners/)
* [Executive Insights](https://www.nvidia.com/en-us/executive-insights/)
* [Startups and VCs](https://www.nvidia.com/en-us/startups/)
* [Documentation](https://docs.nvidia.com/)
* [Technical Training](https://www.nvidia.com/en-us/learn/organizations/)
* [Professional Services for Data Science](https://www.nvidia.com/en-us/support/enterprise/advisory-services/)
Follow NVIDIA
[](https://www.facebook.com/NVIDIA "")[](https://www.instagram.com/nvidia/?hl=en)[](https://www.linkedin.com/company/nvidia/)[](https://twitter.com/nvidia "")[](https://www.youtube.com/user/nvidia)
[United States](https://www.nvidia.com/en-us/location-selector/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [Your Privacy Choices](https://www.nvidia.com/en-us/about-nvidia/privacy-center/)
* [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/)
* [Accessibility](https://www.nvidia.com/en-us/about-nvidia/accessibility/)
* [Corporate Policies](https://www.nvidia.com/en-us/about-nvidia/company-policies/)
* [Product Security](https://www.nvidia.com/en-us/product-security/)
* [Contact](https://www.nvidia.com/en-us/contact/)
Copyright © 2026 NVIDIA Corporation
Select Location
The Americas
* [Argentina](https://www.nvidia.com/es-la/ "Argentina")
* [Brasil (Brazil)](https://www.nvidia.com/pt-br/ "Brasil (Brazil)")
* [Canada](https://www.nvidia.com/en-us/ "Canada")
* [Chile](https://www.nvidia.com/es-la/ "Chile")
* [Colombia](https://www.nvidia.com/es-la/ "Colombia")
* [México (Mexico)](https://www.nvidia.com/es-la/ "México (Mexico)")
* [Peru](https://www.nvidia.com/es-la/ "Peru")
* [United States](https://www.nvidia.com/en-us/ "United States")
Europe
* [België (Belgium)](https://www.nvidia.com/nl-nl/ "België (Belgium)")
* [Belgique (Belgium)](https://www.nvidia.com/fr-be/ "Belgique (Belgium)")
* [Česká Republika (Czech Republic)](https://www.nvidia.com/cs-cz/ "Česká Republika (Czech Republic)")
* [Danmark (Denmark)](https://www.nvidia.com/da-dk/ "Danmark (Denmark)")
* [Deutschland (Germany)](https://www.nvidia.com/de-de/ "Deutschland (Germany)")
* [España (Spain)](https://www.nvidia.com/es-es/ "España (Spain)")
* [France](https://www.nvidia.com/fr-fr/ "France")
* [Italia (Italy)](https://www.nvidia.com/it-it/ "Italia (Italy)")
* [Nederland (Netherlands)](https://www.nvidia.com/nl-nl/ "Nederland (Netherlands)")
* [Norge (Norway)](https://www.nvidia.com/nb-no/ "Norge (Norway)")
* [Österreich (Austria)](https://www.nvidia.com/de-at/ "Österreich (Austria)")
* [Polska (Poland)](https://www.nvidia.com/pl-pl/ "Polska (Poland)")
* [România (Romania)](https://www.nvidia.com/ro-ro/ "România (Romania)")
* [Suomi (Finland)](https://www.nvidia.com/fi-fi/ "Suomi (Finland)")
* [Sverige (Sweden)](https://www.nvidia.com/sv-se/ "Sverige (Sweden)")
* [Türkiye (Turkey)](https://www.nvidia.com/tr-tr/ "Türkiye (Turkey)")
* [United Kingdom](https://www.nvidia.com/en-gb/ "United Kingdom")
* [Rest of Europe](https://www.nvidia.com/en-eu/ "Rest of Europe")
Asia
* [Australia](https://www.nvidia.com/en-au/ "Australia")
* [中国大陆 (Mainland China)](https://www.nvidia.com/zh-cn/ "中国大陆 (Mainland China)")
* [India](https://www.nvidia.com/en-in/ "India")
* [日本 (Japan)](https://www.nvidia.com/ja-jp/ "日本 (Japan)")
* [대한민국 (South Korea)](https://www.nvidia.com/ko-kr/ "대한민국 (South Korea)")
* [Singapore](https://www.nvidia.com/en-sg/ "Singapore")
* [台灣 (Taiwan)](https://www.nvidia.com/zh-tw/ "台灣 (Taiwan)")
Middle East
* [Middle East](https://www.nvidia.com/en-me/ "Middle East")
NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking "Accept All", you consent to our use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/). You can manage your cookie settings by clicking on "Manage Settings." By continuing to use this site or by clicking one of the buttons below, you agree to our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. To opt out of non-cookie personal information "sales" / "sharing" for targeted advertising purposes, please visit the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies. You can manage these settings in the NVIDIA [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
Manage Settings
Reject Optional Accept All
Cookie Settings
We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.
Under certain privacy laws, you have the right to direct us not to "sell" or "share" your personal information for targeted advertising. To opt-out of the "sale" and "sharing" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the "sale" and "sharing" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/).
Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking "Save and Accept" or "Decline All" at the bottom, you consent to the use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) in accordance with your settings and accept our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). For more information about our privacy practices, please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/).
Required Cookies
Always Active
These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.
Cookies Details
Performance Cookies
- [x] Performance Cookies
These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.
Cookies Details
Personalization Cookies
- [x] Personalization Cookies
These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.
Cookies Details
Advertising Cookies
- [x] Advertising Cookies
These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.
Cookies Details
Cookie List
Clear
* - [x] checkbox label label
Apply Cancel
Consent Leg.Interest
- [x] checkbox label label
- [x] checkbox label label
- [x] checkbox label label
Decline All Save and Accept
[](https://www.onetrust.com/solutions/consent-and-preferences/)
Copy debug info